Flex: A flowlet-level load balancing based on load-adaptive timeout in DCN

Load balancing mechanisms spread traffic across multiple paths to achieve high throughput and low latency in data center networks. Flowlet switching has been proven to be a powerful technique for fine-grained load balancing. Existing flowlet-level schemes detect flowlets based on a fixed timeout. However, the fixed timeout is configured empirically: setting it to an excessively large value leads to fewer flowlet opportunities and low link utilization, while setting it to an insufficient value will risk severe problematic disorder. To cope with this problem, we propose Flex, a flowlet-level load balancing scheme based on load-adaptive flowlet timeout. Flex maintains a separate flowlet timeout for each flow, and dynamically updates the timeout value during the transmission of the flow to adapt to changes in path condition. The flow is split into flowlets at the host based on the adaptive timeout. To pass the flowlet results to the switch, Flex marks the adjacent flowlets of the same flow via one bit of the reserved field in TCP header. Experimental results demonstrate that Flex can achieve performance close to that of LetFlow in the baseline topology, and Flex performs particularly well when handling asymmetry and bursty traffic.